Extract the same field from every document

Best when you need the same value out of a stack of same-layout documents — invoices, forms, receipts. Draw a rectangle once on the first to mark what you want, and pull that field from every document into one clean table — with position — that you can download as JSON.

Drop documents to extract from

JPG, PNG, WEBP or PDF — up to 20 MB each; PDF pages become one document each. Add up to 10 documents (same layout); the first is the template you draw regions on, and each document uses one of your monthly extractions. Files are visible only to your account and auto-deleted after 30 days.

How it will work

From a stack of documents to a clean table of values in three steps.

1

Upload

Add a batch of same-layout documents — images or PDFs. Each PDF page is handled as its own document.

2

Mark a region

Draw a rectangle over one document where the value sits and give it a label — the same region applies to every file.

3

Extract

Get every document's value from that region in one table — with position — and download it as JSON.

Working with a batch of same-layout documents

When marking a region beats both typing values by hand and describing a schema to a model.

What this is for

One task, specifically: you have many documents that came off the same template — invoices from one supplier, receipts on one form, shipping labels, application forms — and you need the same value out of each. General OCR gives you every word on every page and leaves you to find the one you wanted, once per document. This tool asks you to mark the value once, on one document, and pulls the corresponding region from all the others into a table.

Why position works as the signal

OCR returns text with coordinates, and most tools discard them. Keeping them turns "the total on this invoice" from a semantic question into a geometric one: no model needs to understand invoices, it just needs to know where you pointed. Regions are stored relative to the page rather than in pixels, so a rectangle drawn on a 1240-pixel scan still means the same thing on a 3024-pixel photo of the same form.

Where it stops working

  • Variable-length content shifts what follows. An invoice with three line items and one with thirty do not put the total in the same place. Anything below a growing table moves with it.
  • Photographs are not scans. Relative coordinates absorb scale, not skew or perspective. A page shot by hand at an angle is the hardest input here.
  • Template changes are silent. If a supplier redesigns their form, the stored region quietly points at the field next door — and returns a confidently extracted wrong value. This is why the output is a table you scan rather than a pipeline that files numbers away unseen.

PDFs

PDF pages are rendered to images and run through the same OCR path as everything else, rather than reading an embedded text layer. That is deliberate: scanned documents in a PDF wrapper have no text layer at all, and generated PDFs often have one whose reading order does not match the visual layout — two-column documents especially. Since this whole approach depends on where things are, a text layer with the wrong geometry is worse than none. The PDF renderer only loads when you actually drop a PDF.

Getting a clean run

  • Draw the region a little larger than the value. Extra whitespace costs nothing; a tight box is fragile against small shifts.
  • Pick a typical document as the template — not the shortest or the longest one, since the region inherits its layout.
  • Keep resolution up. Extraction quality is bounded by OCR quality, which is bounded by pixels per character.
  • Scan the output column before using it. One value of the wrong kind in a list of forty is obvious in seconds — that check is the safeguard against a template change.

What you get back

A table with one row per document: the extracted value, the region it came from, its position, and the source filename. Downloadable as JSON, so it can feed a spreadsheet or a script. Documents run as independent jobs, so one corrupt file fails on its own rather than stopping the batch, and each has a five-minute deadline so nothing hangs indefinitely.

Languages and privacy

The same language support as our other OCR tools: Chinese, Japanese, and 40-plus Latin-script languages. Uploaded documents are deleted immediately after processing — see the Privacy Policy — and if the documents are in a language you cannot read, Image Translator keeps the layout you would navigate by.

Further reading

Letting users fix what OCR missedWhy a correction box beats a better model, and what the corrections tell us.Pulling the same field out of a stack of documentsUsing position as the signal when every page has the same layout.